AI Safety Gridworlds by J Leike et al on Arxiv Vanity 28 November 2017 - The Sentient Robot

The folks at DeepMind are seeking to contribute to AI safety. They have designed a 2D testing environment for algorithms. The environment does not purport to cover all possible AI safety problems. For example, interpretability, multi-agent and formal verification safety problems are not covered. But a decent number is covered.